FinGPT: Democratizing Internet-scale Data for Financial Large Language Models

In one sentence

Summary

The paper addresses the critical gap in accessible, high-quality financial data for training Large Language Models (LLMs), noting that existing models like BloombergGPT are closed-source and prohibitively expensive to train from scratch. The authors argue that general-purpose LLMs underperform in finance due to domain-specific disparities and the time-sensitive nature of market data, necessitating a data-centric approach that democratizes access to internet-scale financial information. This positioning highlights the need for an open, reproducible infrastructure that supports continuous model adaptation to evolving market conditions. To solve this, the authors propose FinGPT, a four-layer framework comprising data sourcing, curation, LLM adaptation, and application. The data layer aggregates real-time information from 34 diverse sources, including news, social media, SEC filings, and academic datasets, via unified APIs. The curation pipeline automates cleaning, filtering, and deduplication to handle low signal-to-noise ratios. For model adaptation, the framework employs Low-rank Adaptation (LoRA/QLoRA) for cost-efficient fine-tuning and introduces Reinforcement Learning with Stock Prices (RLSP), which uses automated market feedback (stock price changes) as labels instead of costly human annotations. Experiments demonstrate the framework's utility in robo-advisory, sentiment analysis for quantitative trading, and low-code development. Empirical results show that FinGPT significantly outperforms base models like LLaMA in financial sentiment classification, particularly when excluding neutral labels, with an 198% improvement in accuracy for positive/negative classification. In simulated quantitative trading, FinGPT achieved a 9.5% average cumulative return compared to -0.1% for the baseline, validating the effectiveness of RLSP and high-quality curated data. The paper emphasizes privacy and openness, allowing users to train local LoRA weights without data leakage. Limitations include the current focus on US and CN markets, the simplicity of the 2% threshold for RLSP labeling, and the need for future work on longer context windows, RAG integration, and bias mitigation.

Research problem

Existing financial LLMs suffer from limited access to high-quality, real-time data and high training costs, with proprietary models like BloombergGPT being closed-source. General-purpose LLMs also underperform in finance due to domain-specific nuances and the time-sensitive nature of market data. The paper addresses the challenge of democratizing access to internet-scale financial data and enabling efficient, continuous adaptation of LLMs to dynamic market conditions without relying on expensive human-annotated feedback.

Core contributions

• Introduction of FinGPT, an open-source, data-centric framework for developing financial LLMs.
• Development of an automated data curation pipeline integrating 34+ diverse financial data sources (news, social media, filings) with real-time APIs.
• Proposal of Reinforcement Learning with Stock Prices (RLSP), a novel method for fine-tuning LLMs using automated market feedback instead of human annotations.
• Implementation of lightweight adaptation using LoRA and QLoRA to reduce training costs to approximately $262.
• Empirical validation of FinGPT in three applications: robo-advisory, quantitative trading sentiment analysis, and low-code development.

Method

• Automated data collection from diverse internet sources via unified APIs.
• Real-time data curation pipeline including cleaning, filtering, deduplication, and tokenization.
• Reinforcement Learning with Stock Prices (RLSP) for automated labeling based on 2% stock price thresholds.
• Parameter-efficient fine-tuning using Low-rank Adaptation (LoRA) and Quantized LoRA (QLoRA).
• Sentiment analysis classification into positive, negative, and neutral categories based on news text.
• Simulated trading strategy: buying/selling based on sentiment signals with a 5-day holding period.

Results & findings

• FinGPT achieves 0.481 overall accuracy in sentiment prediction, a 6.8% improvement over LLaMA.
• Accuracy for positive/negative classification (excluding neutral) improves by 198.4% (0.063 to 0.188) compared to LLaMA.
• Macro-F1 score for positive/negative classification improves by 103.4% (0.035 to 0.071).
• Simulated trading using FinGPT signals yields a 9.5% average cumulative return, compared to -0.1% for LLaMA.
• Training cost for fine-tuning is reduced to approximately $262 using LoRA/QLoRA, compared to millions for full pre-training.
• The RLSP method effectively leverages market feedback, eliminating the need for expensive human annotation for fine-tuning labels.

Data & experiments

• Data sources include 19 news outlets, 8 social media platforms, 3 filing sources (SEC, etc.), and 4 academic datasets (FPB, FiQA SA, etc.).
• Experiments on sentiment analysis for quantitative trading using S&P 500 constituents and Yahoo Finance price data.
• Comparison of FinGPT against base LLaMA and other models on financial sentiment classification benchmarks.
• Evaluation of trading performance via simulated trading experiments measuring average cumulative return rate (CRR).
• Assessment of fine-tuning efficiency and cost reduction using parameter-efficient methods like LoRA.

Limitations & caveats

• Current focus is limited to US and CN stock markets; other markets are not covered.
• The RLSP labeling uses a simple 2% threshold, which may not capture all market nuances.
• The framework relies on publicly available data, potentially missing proprietary or high-quality private data.
• Sentiment analysis results tend to lean towards neutrality, which may limit signal strength for trading.
• The paper acknowledges potential bias and fairness issues that require future investigation.
• Context window limitations may restrict performance on long-form financial document analysis.

Curator notes

Discovery status: round2_promoted. Source links: A Comprehensive Review of Generative AI in Finance; CFGPT: Chinese Financial Assistant with Large Language Model; DISC-FinLLM: A Chinese Financial Large Language Model based on Multiple Experts Fine-tuning; Enhancing Financial Sentiment Analysis via Retrieval Augmented Large Language Models; Evaluating Financial Intelligence in Large Language Models: Benchmarking SuperInvesting AI with LLM Engines; FinLlama: Financial Sentiment Classification for Algorithmic Trading Applications; InvestLM: A Lar